MFGC-Net: Bridging and Fusing Multiscale Features and Global Contexts for Multitask Sea Ice Fine Segmentation

Sea ice segmentation from synthetic aperture radar (SAR) imagery is a key task in polar sea ice monitoring, which is crucial for global climate prediction and polar route planning. However, the existing sea ice segmentation algorithms for SAR images often fail to consider long-range contextual dependencies when capturing multiscale features, resulting in an inability to fully exploit multiscale global contextual information. To address this limitation, we proposed a novel encoder–decoder structure network for multitask sea ice segmentation. Initially, a cross-scale interaction module was constructed in the encoder that utilizes cross attention to seamlessly capture multiscale features, effectively bridging the semantic information gap between different layers. Subsequently, a context transformer block based on efficient multihead self-attention was developed to model remote dependencies across spatial and channel dimensions, thereby enhancing the extraction of multiscale global contextual information. Furthermore, a channel patch module was introduced that allows for the strategic refinement of differential features to emphasize changing areas and suppress artifacts. In the final stages, a refined multiscale feature fusion module was embedded in the decoder to strategically integrate the feature maps generated, thus iteratively merging layered features for enhanced segmentation. Our experiments on the AI4Arctic Sea Ice Challenge Dataset show that MFGC-Net achieves outstanding performance in multitask sea ice segmentation compared with current state-of-the-art methods, as demonstrated by both quantitative and qualitative results.